wasnaga/punk-records-research-kernel-v0.1

Punk Records Research — Agent Kernel v0.1

Research artifact — not a chat LLM.

Punk Records — lawful agent memory as kernel state transitions (One Piece punk records metaphor: immutable logs that outlive any speaker).

Layer	Artifact	Role
Kernel law	policies/*.json + KERNEL.md	8-opcode invariants, gas, overflow
Encoder E6	encoder_e6_best.pt	Learned opcodes + slot keys + values
NL parser E7	encoder_e7_best.pt	Learned intent/slot + percept value bootstrap
Renderer E3	renderer_e3_best.pt	Untrusted text from storage

Source code: github.com/umbecanessa/punk-records-research

---

Results (v0.1)

Metric	Score
Mean reward (A–F)	0.991
Query accuracy	1.0 all stages
Render fidelity	1.0 full stack
Value exact match	1.0
E7 NL parse (intent / slot / value)	1.0 / 1.0 / 1.0

Frozen reports in eval/. Reproduce with the GitHub repo + these weights.

---

Download

pip install -U "huggingface_hub[cli]"
huggingface-cli download wasnaga/punk-records-research-kernel-v0.1 --local-dir ./kernel-v0.1

File	Description
encoder_e6_best.pt	Release encoder (~460 KB)
encoder_e7_best.pt	NL event parser (~510 KB)
encoder_e10a_best.pt	Open-phrasing NL parser v2, 96-char (~7 MB)
encoder_e9a_best.pt	Transformer NL front (~3 MB)
renderer_e9b_best.pt	Transformer renderer (~3 MB)
renderer_e3_best.pt	Byte renderer (~5 MB)
stack.json	Bundle manifest
policies/	Kernel policy JSON
eval/	Training eval reports
KERNEL.md	Opcode ABI summary

---

Architecture

structured event + machine state
        ↓
  encoder E6 (MLP)
        ↓
  opcode trace (OBS → PUT → SEAL → … → GET → RENDER)
        ↓
  deterministic kernel
        ↓
  storage (+ cold overflow)
        ↓
  renderer E3 → answer text

Values at inference come from OBS percept bytes in features, not chat payload strings.

---

Metrics

• query_accuracy — storage read matches world truth
• revert_rate — invalid ops rejected by kernel
• render_fidelity — template match (renderer untrusted)
• reward — query_accuracy − 0.5 × revert_rate

---

Citation

@software{punk_records_research_kernel_v01,
  title  = {Punk Records Research — Agent Kernel v0.1},
  year   = {2026},
  url    = {https://huggingface.co/wasnaga/punk-records-research-kernel-v0.1},
  note   = {8-opcode session state machine; structured-event training.}
}

License

Apache-2.0